Collinearity because of technical problem in Fixed Effects Panel Regression

[Salman Mallick]

Hi,

I am trying to look at the differences in income between African-Americans and White Americans. I am using panel data, and have run a fixed effects regression, which is the following:

xtreg linc i.highest_degree#bw i.bw#c.lhrswrk i.occup#bw ib2.urban_rural#bw i.marriage#bw i.bw#c.lwkswk_job i.census_region#bw, fe baselevel

However, I am also running into a problem with collinearity, although it is strange in that it seems almost systematic:
At the end of every interaction with a categorical variable, I have one value which is always omitted; that too always the very last category in the categorical variables. It doesn't happen for the continuous ones.

note: 1.educ#1.bw omitted because of collinearity
note: 10.occup#1.bw omitted because of collinearity
note: 2b.urban_rural#1.bw omitted because of collinearity
note: 5.marriage#1.bw omitted because of collinearity
note: 4.census_region#1.bw omitted because of collinearity

         marriage#bw |
Never married#White  |          0  (base)
Never married#Black  |  -.5309271   .0851924    -2.02   0.043    -.3394337   -.0054707
      Married#White  |   .8552037   .0204159     7.87   0.000     .1206852    .2007177
      Married#Black  |   .4691787   .0831313     0.69   0.489    -.1054266    .2204569
Separ, Divorc#White  |   .0947397   .0391489     4.68   0.000     .1065997    .2600673
Separ, Divorc#Black  |          0  (omitted)

                     |
     bw#c.lwkswk_job |
              White  |   .1013588    .004665    21.73   0.000     .0922152    .1105023
              Black  |   .1059799   .0079389    13.35   0.000     .0904193    .1215405

Is this a so-called 'technical problem', a problem with my formulation, or is it actually due to collinearity that STATA is finding in the data? Thank you very much.

[Carlo Lazzaro]

Salman:
1) some omitted variable might be due to the -fe- estimator behaviour (ie, wiping out every regressors that is time-invariant);
2) other collinearity issues among the remaining predictors can exist.
Eventually, I would not sponsor the way you build up interactions, as you seemingly kicked out from the set of predictors the so called conditional main effect of the predictors included in interactions.

[Salman Mallick]

Hi Carlo,

Thank you for your advice.

Regards,
Salman